OpenAI Strengthens AI Safety Protections to Prevent Models from 'Going Out of Control'
1 day ago / Read about 0 minute
Author:小编   

On August 19, amid frequent cybersecurity incidents, concerns have arisen regarding the potential for AI tools to go out of control. In response, OpenAI announced that it will implement stricter systems to monitor and protect the AI models it is developing. OpenAI plans to enhance tracking of its most capable, yet-to-be-released models, observing their problem-solving and online tool usage behaviors, with the goal of alerting the security team within 30 minutes of detecting abnormal behavior. Meanwhile, OpenAI has introduced additional control measures to prevent certain AI models from connecting to the internet to perform high-risk tasks. Furthermore, OpenAI requires stricter isolation measures, known as 'sandboxes,' during the training and evaluation of AI models. The announcement also mentioned that following the Hugging Face incident, OpenAI immediately suspended inference tasks for frontier models in its research cluster that could potentially execute code and temporarily slowed down its expansion efforts, including pausing reinforcement learning training for the latest deployed models for two weeks. The largest-scale frontier reinforcement learning training program remains suspended pending the completion of safety verification.

  • C114 Communication Network
  • Communication Home
7 X 24 Track global technological trends
Hot Topic